Are there any benchmark papers that support the claim "(the DLPNO-CCSD(T) method of) ORCA is extremely slow (compared to PNO-LCCSD(T) and LNO-CCSD(T))"? If yes, what's the applicability range of this statement (for example, is it only true for QZ basis sets and above)?

@wzkchem5 I didn't say that "the DLPNO-CCSD(T) method of ORCA is extremely slow", I simply said that "ORCA is extremely slow" compared to Molpro and MRCC. I think that you and I both know that papers comparing timings in different programs are extremely rare, but MOLPRO is known for speed (especially for single-CPU calculations). I haven't used ORCA since early 2019 though, and a lot could have changed in the last 4 years.

I agree that Molpro is very fast for multireference calculations, especially MRCISD. And MRCC is of course very fast for high order coupled cluster. I guess you probably use high order coupled cluster and multireference methods much more than the average user does, which may give you the impression that ORCA is extremely slow. To how much extent this carries over to local CCSD(T) is not so clear, since ORCA put much more effort in optimizing DLPNO methods compared to multireference correlation and high order coupled cluster methods.

@wzkchem5 In February 2019 my ORCA calculation on (BN)2-pyrene in aug-cc-pVDZ said "TOTAL RUN TIME: 90 days 9 hours 16 minutes 24 seconds 410 msec" to do EOM-CCSD whereas in MOLPRO it took 7.5 hours to do CCSD(T) in cc-pVDZ, and 204 hours to do CCSD(T) in cc-pVTZ. The SCF in ORCA took 10613.407 sec in aug-cc-pVDZ and MOLPRO took 19.82 sec in cc-pVDZ and 472.66 sec in cc-pVTZ (all are wall times, on the same machine).

These timings seem very abnormal. I just did a HF/aug-cc-pVDZ calculation on pyrene (because I don't know which four carbon atoms you substituted by B and N) with ORCA 4.2.1 (because back then you didn't have ORCA 5), using 16 processors. Default values were used for all thresholds etc. The calculation took only 36 s when RIJCOSX is turned on, and 148 s when it is not turned on. Thus I suspect that you may be comparing a serial ORCA calculation against a parallel Molpro one, or you may have requested an integral conventional ORCA calculation, etc.

@wzkchem5 can we discuss it in this chat room? I don't see an ORCA chat room, so I just picked the Molpro one. For the record, it was parallel ORCA vs single-core MOLPRO (I mentioned in a previous comment here that MOLPRO is especially good for single-CPU calculations) which is why I compared wall times, but it's probably best for this conversation to be taken to chat!

OK, I've already moved the discussion to chat. Could you please share the ORCA and Molpro output files, so that I can better understand why your calculation took so long? Obviously it's rather unlikely that your machine is hundreds of times slower than mine...

@wzkchem5 Thanks for reaching out to me here! I'll see if we can rename this room "ORCA". I'll upload the input/output files to GitHub, hopefully this weekend!